Remote Workflow

This tutorial walks through a practical laptop-to-server workflow using the current Remote Viewer stack.

Scenario

You train on a remote GPU machine but want to browse runs from your local machine without syncing the whole storage directory.

1. Prepare the remote machine

Install Runicorn into the Python environment you actually train with:

conda activate train-env
pip install runicorn

Write runs into a stable remote storage root:

import runicorn as rn

run = rn.init(
    path="remote-demo/resnet50",
    storage="/data/runicorn",
    alias="seed-42",
)

run.log_config(extra={"machine": "gpu-server-01", "epochs": 3})

latest_val_acc = 0.0

for epoch in range(1, 4):
    for _ in range(5):
        run.log({
            "train_loss": 1.0 / epoch,
            "val_acc": latest_val_acc,
        }, stage=f"epoch_{epoch}")

    latest_val_acc = 0.70 + 0.05 * epoch

run.summary({
    "final_val_acc": latest_val_acc,
    "remote_root": "/data/runicorn",
})

run.finish()

2. Start local browsing

On your local machine:

runicorn viewer

Open the Remote page.

3. Add the server and connect

Create a saved server entry, authenticate, confirm the host key if prompted, and choose the remote environment.

4. Start the remote session

After the session starts, the remote runs appear through the forwarded viewer just like local runs.

Watch for statuses such as:

running
reconnecting
degraded

5. Monitor the training job

Use the normal UI once connected:

experiments page
run detail charts
logs
images
assets

6. Stop cleanly

Use the Stop action in the Remote page when you are done. Runicorn 0.7.0 also cleans up the SSH connection when that session is the last one using it.